The Gate Before the Model

Published Sep 19, 2026

Sequel to Where Human Judgment Still Wins.

In Where Human Judgment Still Wins, the comfort blanket was "put a human in the loop." This time the comfort blanket is a different phrase with the same smell: "we’re doing AI." A backlog of pilots is not a strategy. A model without an outcome is a science fair. And a company that cannot say what should enter the AI pipeline — or what "done" means before the tool runs — is not short on technology. It is short on gates. If your AI strategy fits on a slide titled Transformation and nowhere else, you do not have a strategy. You have typography.

I am COO of BSC Analytics. I run the operating company. My job is not to collect demos like commemorative pins. It is to keep delivery honest when tools meet real work: incomplete inputs, contested ownership, and people who have to live with the result on Tuesday. AI raises the floor on speed and pattern. Humans still decide what gets on the floor in the first place — which, inconveniently, requires saying no.

In 2025, MIT’s NANDA initiative published The GenAI Divide: State of AI in Business 2025 — interviews with 150 leaders, a survey of 350 employees, and analysis of about 300 public deployments. The headline that traveled was brutal: about 95% of enterprise GenAI efforts showed no measurable P&L impact. The useful part is why. Researchers pointed less at model quality and more at a "learning gap" — brittle workflows, weak integration into day-to-day operations, tools that do not adapt to how the work actually runs. In other words: the failure mode was organizational, not cinematic. Nobody loses because the robot forgot its lines. They lose because the company never wrote a script the operating system could run. A RAND report on why AI projects fail landed in a similar place from practitioner interviews. One of the dominant root causes was not "the algorithm was dumb." It was business leadership misunderstanding how to set the project on a pathway to success — objectives that never translated cleanly into what the model should optimize, or success metrics that measured the wrong thing. If you ask for the price that sells the most units when what you needed was the price that protects margin, the model can be excellent and still be useless. That is not an anti-AI story. It is an intake story — and intake is where ego usually arrives before clarity.

Governance is not a committee that meets after the pilot already has a name, a vendor, and a heroic Slack channel. Governance is the set of human decisions that happen before the tool earns a calendar slot. What gets into the pipeline? Not every idea with an API and a hopeful grin — work that has a named owner, a real workflow, and a problem statement that survives contact with operations. How is the outcome defined before tools run? Baseline, target, time window — in language a delivery team can execute, not a slide that says "transform" in a font that looks expensive. Who decides priorities when everyone wants to be next? Someone has to say yes to one pain point and no to twelve fashionable ones. That someone will briefly be less popular than the model. Where does judgment stay in the loop? Not as a rubber stamp at the end. As the authority that can stop the work, redefine done, or kill a pilot that cannot fail cleanly. AI raises the floor under execution. Imagination, judgment, relationships, and taste still decide which problems deserve that execution — and what "better" means when the output is plausible and wrong.

From an ops seat, the intake gate is more important than the applause at the demo. Applause is cheap. Baselines are not. Before anything enters the AI pipeline, I want answers that sound boring on purpose. What unit of work starts and finishes? What number moves, from what baseline, by when? Who owns the outcome when the vendor leaves the room? What is the kill criteria if the pilot only produces theater? If those answers are fog, the model is not late. The request is early. Prioritization is not ranking slogans. It is sequencing scarce attention. The companies that cross the GenAI Divide in the MIT framing tend to pick a real pain point, execute against business metrics, and hold the work accountable to the operating system — not to the excitement of the tool. The ones stuck on the wrong side keep funding static experiments that never learn the workflow, then wonder why the scoreboard looks like a museum of almost.

The most expensive AI failure is not a bad model. It is a well-run model pointed at an undefined outcome. Judgment is the gate. Intake is where judgment shows up as courage: the willingness to keep half the backlog out so the other half can become real. Saying no is not anti-innovation. It is how innovation gets a calendar.

AI will keep getting better at raising the floor. The companies that get value will not be the ones with the longest pilot list. They will be the ones where humans still set the gates — what enters, what "done" means, what comes first, and what never should have been in the queue.